← All Working Papers & Datasets

Empirical Working Paper · AI Systems Security & Software Supply Chains

Supply-Chain Security in Open-Source AI Agent Ecosystems: Empirical Evidence and a Five-Dimension Audit Framework Across 105,000+ Indexed Agent Skills and Model Context Protocol (MCP) Servers

Jason Zhu (祝彦森) · BRICKAI (GoSailGlobal LLC) / JasonZhu.AI

ORCID: 0009-0000-2894-6285Date: 2026-11-217 pagesACM CCS: Security and privacy → Software and application security

Abstract

Autonomous coding agents—including Claude Code, OpenAI Codex, and Cursor—increasingly rely on community-authored skills and Model Context Protocol (MCP) servers installed directly from GitHub with full local shell and credential privileges. Drawing on Agent Skills Hub's open census (105,000+ entries in the benchmark audit snapshot; 130,173+ graded catalog records archived at CERN Zenodo and Wikidata Q141735092), this paper reveals an acute visibility gap: 97.8% (102,690 of 105,000+) indexed skills had never undergone any published security review. Across 2,300+ deeply audited skills, 94.4% passed cleanly while 5.6% triggered caution, unsafe, or rejection flags; across 21,582 popular rule-graded repositories, approximately 1 in 30 receives an UNSAFE or REJECT rating. We formalize a five-dimension supply-chain audit framework (Code Layer, Credential Handling, Vendor Provenance, Dependency Supply Chain, and Operational Risk) and demonstrate through three real-world case studies why GitHub stars fail as a proxy for security and engineering rigor.

AI Agent SecurityModel Context ProtocolMCP ServersClaude SkillsSoftware Supply ChainStatic AnalysisPrompt InjectionAgentSkillsHub

Supplementary Open Datasets (CC-BY-4.0)

BibTeX Citation

@techreport{zhu2026agentskillssecurity,
  author      = {Zhu, Jason (Yansen Zhu)},
  title       = {Supply-Chain Security in Open-Source AI Agent Ecosystems: Empirical Evidence and a Five-Dimension Audit Framework Across 105,000+ Indexed Agent Skills and Model Context Protocol (MCP) Servers},
  institution = {Agent Skills Hub (BRICKAI / GoSailGlobal LLC) and JasonZhu.AI},
  year        = {2026},
  month       = {11},
  type        = {Working Paper},
  doi         = {10.5281/zenodo.24420920},
  url         = {https://papers.ssrn.com/sol3/papers.cfm?abstract_id=7737258},
  note        = {ORCID: 0009-0000-2894-6285; SSRN Abstract ID: 7737258; Wikidata: Q141735092}
}